fix: v2.5.1 security/ML/reliability patch

Security:
- H1 DNS rebinding TOCTOU: validate_endpoint returns (endpoint, resolved_ips);
  create_pinned_session builds an aiohttp session with a custom resolver that
  returns only the pre-validated IPs, closing the re-resolve gap.
- H2/M3 Shared session cookie pollution: integration no longer uses HA shared
  clientsession; isolated session with DummyCookieJar prevents cross-integration
  cookie leaks.
- Cloud metadata / link-local block added to allow_local_network mode to
  prevent IMDS exfiltration on cloud VMs.
- L3 hard cap extended to history.async_get_history (not only service schema).

ML/LLM correctness:
- _is_openai_reasoning_model uses regex with gpt-5-chat* blacklist and handles
  OpenRouter-style openai/ prefix. Future o5/gpt-6 auto-recognized.
- reasoning_effort=minimal for gpt-5 family, low for o-series.
- Gemini 2.5 Pro gets thinking_budget=128 (Pro rejects 0, flash accepts 0).
- Anthropic extracts first type=text block instead of hardcoded content[0].
- /no_think dedup uses word-boundary regex instead of substring.
- DeepSeek-reasoner detected: skips /no_think; preserves reasoning_content.

Reliability:
- Exception chaining added to Gemini-block re-raises.
- Top-level imports for json/re/hashlib (no more lazy stdlib imports).
- allow_local_network logged at INFO, not WARNING on every setup.

Code style:
- PEP 604 type hints across all .py files (dict/list/| None).

No breaking changes for users. Internal API: validate_endpoint return type
changed from str to tuple[str, list[str]].

custom_components/ha_text_ai/const.py

@@ -29,7 +29,7 @@ API_PROVIDERS: Final = [
     API_PROVIDER_GEMINI
 ]
 
-VERSION: Final = "2.5.0"
+VERSION: Final = "2.5.1"
 
 # Default endpoints
 DEFAULT_OPENAI_ENDPOINT: Final = "https://api.openai.com/v1"